This is the main file for model interpretation and analysis. Other files (superset-model-analysis and stories-from-superset-models) contain bits and pieces of thoughts and plots but all is consolidated here.
Having standardized the socio-demographic variables within each election year, and by using consistent model specification across elections, we can directly compare the estimated effects of covariates across elections and hence draw conclusions on how the two-party preferred vote has been affected by socio-demographics.
Five of the ten variables in the model are not included in an interaction, so we can interpret their marginal effects directly. To illustrate these effects, we plot 95% confidence bands for each coefficient over the six elections.
For a variable that is involved in an interaction, we can interpret its main effect as being the marginal effect of a one standard deviation increase in that variable, when any variables it is interacted with assume their mean value. This is because the mean value of a standardized variable corresponds with a value of zero. We interpret the interactions using visualizations from the visreg package, which allow us to tease out the effects of each variable, conditioning on a range of values for the variable it is interacted with.
The following section details the interpretation of our models across the six elections. We begin the interpretation of each variable by outlining the effects we expect to observe on the basis of scatterplots with the response, and any other intuition relating to a particular policy proposed by either party in that election. These have no bearing on the relationships determined by the model, but rather provide additional insight into the problem.
Education policy is typically a key policy issue in the weeks leading up to a federal election. The Labor party, more often than not, campaign with policies that boost funding for education, and the coalition will usually propose alternatives with less funding or aim to match particular objectives - for example, the “Better Schools Plan” in 2013. It would be reasonable to expect that electorates with higher student populations would be inclined to support Labor at every election, as we observe in the scatterplots. However, after accounting for other socio-demographic variables, we find that student populations were insignificant in electorate party preference for all but 2010 - in which higher student populations had a positive relationship with the Liberal party.
## Warning: Ignoring unknown aesthetics: label
## Warning: Removed 9 rows containing non-finite values (stat_smooth).
## Warning: Removed 9 rows containing missing values (geom_point).
Rental prices and mortgage loan repayments do not appear to have a signficant relationship with party preference, based on the scatterplots. We find that in the 2010 election, higher median payment amounts had a highly positive relationship with support for the Liberal party. There is no obvious explanation for this, as housing affordability was not on the political agenda in the 2010 election.
## Warning: Ignoring unknown aesthetics: label
## Warning: Removed 9 rows containing non-finite values (stat_smooth).
## Warning: Removed 9 rows containing missing values (geom_point).
The raw data plots indicate that areas with more immigrants from South-Eastern Europe should have a strong relationship with Labor support. We find that this holds true after adjusting for other socio-demographics across all elections, as the estimated coefficients are significantly negative and do not change much across years.
## Warning: Ignoring unknown aesthetics: label
## Warning: Removed 9 rows containing non-finite values (stat_smooth).
## Warning: Removed 9 rows containing missing values (geom_point).
We anticipated the effects of greater populations from the UK would have a weakly positive (likely insignificant) relationship with support for the Liberal party, which is confirmed by our model. Born_UK is insignificant in every election.
Therefore, we conclude that an electorate’s distribution of birthplace is influential, but that only migrants from particular areas have this influence (e.g. South-Eastern Europeans).
## Warning: Ignoring unknown aesthetics: label
## Warning: Removed 9 rows containing non-finite values (stat_smooth).
## Warning: Removed 9 rows containing missing values (geom_point).
We do not anticipate that rates of atheism should have a significant effect on party preference, as there is no sign of a significant relationship in the scatterplots. From 2010 onwards, we find that higher rates of atheism were aligned with support the Labor party, and in other elections this was insignificant.
## Warning: Ignoring unknown aesthetics: label
## Warning: Removed 9 rows containing non-finite values (stat_smooth).
## Warning: Removed 9 rows containing missing values (geom_point).
## Warning: Ignoring unknown aesthetics: label
## Warning: Removed 9 rows containing non-finite values (stat_smooth).
## Warning: Removed 9 rows containing missing values (geom_point).
## Warning: Ignoring unknown aesthetics: label
## Warning: Removed 9 rows containing non-finite values (stat_smooth).
## Warning: Removed 9 rows containing missing values (geom_point).
In a similar way that we expected Labor’s favourable education policies to influence student populations’ voter preference, we might expect education levels to behave in the same way - although the scatterplots do not indicate much of a relationship. On the contrary, older age profiles (higher median age) appear to have a strong relationship with Liberal support.
We find that the electorates with below-average age profiles and above-average levels of education consistently have higher levels of support for the Labor party. Up until 2016, areas with lower education levels were unaffected by age and supported the Liberal party, whereas in 2016 younger electorates supported the Labor party - even with low levels of education. Overall, it is clear the young electorates with higher education levels have always supported Labor, and older electorates with lower education levels have always supported the Liberal party.
## Warning: Ignoring unknown aesthetics: label
## Warning: Removed 9 rows containing non-finite values (stat_smooth).
## Warning: Removed 9 rows containing missing values (geom_point).
Labor often propose policies that benefit families, and particularly single parent families. The scatterplots indicate that there may be a strong negative relationship between OneParent_House and support for the Liberal party. We find that this is mostly true across levels of education. For those with mean or higher education, the effect of single parent households is strong, with higher incidence of single parent households supporting the Labor party. A pocket of Liberal support appears at below-average levels of both variables.
Two industries of work are included in our model - extractive and office jobs. Extractive includes jobs relating to mining, gas, water, gas, waste, electricity and agriculture, and office jobs include administration, clerical, sales and managerial roles. We expect that higher incidences of these should both have strong positive effects on support for the Liberal party.
## Warning: Ignoring unknown aesthetics: label
## Warning: Removed 9 rows containing non-finite values (stat_smooth).
## Warning: Removed 9 rows containing missing values (geom_point).
## Warning: Ignoring unknown aesthetics: label
## Warning: Removed 9 rows containing non-finite values (stat_smooth).
## Warning: Removed 9 rows containing missing values (geom_point).
Increasing percentages of people employed in either industry, holding the other fixed, have increased support for the Liberal party in all elections. The interaction coefficient is close to zero in all elections aside from 2010, meaning that effects are practically independent. In 2010, having a very high percentage of workers in one industry and low in the other resulted in a zero net-effect on two party preference, whereas in other years the prevailing effect is that of the industry with the high percentage.
Furthermore, from 2010 onwards, the effect of extractive workers appears to be stronger for the same value of office jobs, relative to earlier years. We also see that the fixed effect coefficient jumps in 2010, which may be a direct result of Labor proposing a 40% tax on “super profits” made from ‘the exploitation of Australia’s non-renewable resources’. This tax was heavily opposed by the Liberal party and was expected to have an adverse effect on many of these extractive industries. This Labor policy appears to have had a lasting effect on electorates with high proportion of extractive workers.
Overall, it is clear that both industries have had significant positive associations with the Liberal party across all elections.
Over all years we see that for moderate-to-high proportions of workers in office jobs, increased levels of education are related to increased support for the Labor party, but this effect has varied for low proportion of office jobs. Electorates with low values for both office jobs and education are consistently associated with supporting the Labor party. No electorates exist with high levels of education, and low proportions of office jobs in any election. We also see that high proportions of office jobs and low levels of education are likely to support the Liberal party in every election, whereas high proportions of office jobs and high levels of education are assocated with the Labor party.